官网 Early access 申请地址

研究方向

We give Grok access to search tools and real-time information, but as with all the LLMs trained on next-token prediction, our model can still generate false or contradictory information. We believe that achieving reliable reasoning is the most important research direction to address the limitations of current systems. Here, we would like to highlight a few promising research directions we are most excited about at xAI:

• Scalable oversight with tool assistance. Human feedback is essential. However, providing consistent and accurate feedback can be challenging, especially when dealing with lengthy code or complex reasoning steps. AI can assist with scalable oversight by looking up references from different sources, verifying intermediate steps with external tools, and seeking human feedback when necessary. We aim to make the most effective use of our AI tutors' time with the help of our models.
• Integrating with formal verification for safety, reliability, and grounding. To create AI systems that can reason deeply about the real world, we plan to develop reasoning skills in less ambiguous and more verifiable situations. This allows us to evaluate our systems without human feedback or interaction with the real world. One major immediate goal of this approach is to give formal guarantees for code correctness, especially regarding formally verifiable aspects of AI safety.
• Long-context understanding and retrieval. Training models for efficiently discovering useful knowledge in a particular context are at the heart of producing truly intelligent systems. We are working on methods that can discover and retrieve information whenever it is needed.
• Adversarial robustness. Adversarial examples demonstrate that optimizers can easily exploit vulnerabilities in AI systems, both during training and serving time, causing them to make egregious mistakes. These vulnerabilities are long-standing weaknesses of deep learning models. We are particularly interested in improving the robustness of LLMs, reward models, and monitoring systems.
• Multimodal capabilities. Currently, Grok doesn’t have other senses, such as vision and audio. To better assist users, we will equip Grok with these different senses that can enable broader applications, including real-time interactions and assistance.

We believe that AI holds immense potential for contributing significant scientific and economic value to society, so we will work towards developing reliable safeguards against catastrophic forms of malicious use. We believe in doing our utmost to ensure that AI remains a force for good.